A Study On Non-negative Matrix Factorization And Self-supervised Learning Algorithm For Image Classification

Supervised learning-based network models have achieved excellent results in many real-world tasks,but their dependence on sample labels for training still has significant limitations in practical scenarios.At the same time,complex data often faces problems such as the explosion of dimensions during processing.Existing machine learning algorithms,whether traditional or based on deep neural networks,extract features from data to obtain low-dimensional representations,but these methods cannot guarantee that the extracted features are an approximate representation of the original data or effective feature information for downstream tasks.Among many machine learning algorithms,non-negative matrix factorization can effectively extract features from high-dimensional data,but it still has certain deficiencies,such as not considering the manifold relationship between images and the sparsity of high-dimensional data affecting the results obtained by the algorithm during matrix factorization.Additionally,in the field of deep learning,self-supervised learning is famous for achieving good results without the need for manual labeling of data,but the quality of auxiliary tasks directly determines the results of self-supervised models,and the use of large amounts of unlabeled data requires significant computing costs.This paper proposes two related algorithms for image classification in two different fields,one is non-negative matrix factorization in traditional machine learning,and the other is self-supervised learning in deep neural networks.The main contributions of this paper are as follows:(1)In the field of traditional machine learning,a new image classification algorithm for face recognition based on graph learning regularized discriminative non-negative matrix factorization has been proposed by combining adaptive manifold learning with non-negative matrix factorization.In this algorithm,graph learning and the self-representation matrix of the subspace are first introduced into the construction process of the Laplacian matrix of manifold learning,allowing the relationship between images to be preserved during the matrix factorization process.At the same time,incorporating label information enhances the model’s classification ability.(2)In the field of deep learning,this paper proposes Two Momentum Contrast in Triplet for Unsupervised Visual Representaion Learning(TMCT)by using a ternary network structure to integrate contrastive and non-contrastive self-supervised learning.In this algorithm,different positive sample pairs are first constructed,and the feature representation of images is learned through the contrast between positive and negative samples.Then,the obtained representation is introduced into another space to further learn excellent image features through the contrast between positive sample pairs.(3)In the field of engineering application,this article constructs a corresponding image classification system based on the dual-momentum contrastive learning algorithm.The system consists of three parts: the front-end,back-end,and algorithm modules.In addition,the front-end and back-end jointly build login and upload functions,and the back-end and algorithm jointly build online classification and model training functions.Furthermore,the system uses a simple architecture design to ensure its stability and conducts relevant testing.Ultimately,the entire system can accurately classify input images.